In order to analyze if there is a mediation effect using Baron & Kenny's steps, is it necessary to include the control variables of my model, or is it enough to do the analysis just with the independent variable, the mediator variable and the dependent variable of my interest?
Well it seems this is a test of what should go into a particular path model, which is most often NOT a predictive. model. A reasonable description of the state of the art is given here: https://www.google.com/search?q=terena+mkt-8543-quantitative-marketing-seminar&rlz=1C1CHBF_enUS874US874&oq=terena+mkt-8543-quantitative-marketing-seminar&aqs=chrome..69i57.24028j0j4&sourceid=chrome&ie=UTF-8
However because of recent developments in regression modeling, for example see: Article To Explain or to Predict?
, it might be wise to expect some modification in the future as has been the case with prediction models for example see: https://www.amazon.com/Clinical-Prediction-Models-Development-Validation/dp/3030163989/ref=sr_1_1?crid=3MIGZGH6E366X&keywords=ewout+steyerberg&qid=1577598485&sprefix=steyerb%2Caps%2C254&sr=8-1 For the time being the above ppt presentation is standard and thus we should use the current full model for the path of interest. A helpful link might be: https://www.google.com/search?q=Regression+Models+in+Path+Analysis&rlz=1C1CHBF_enUS874US874&oq=Regression+Models+in+Path+Analysis&aqs=chrome..69i57.23826j0j4&sourceid=chrome&ie=UTF-8D. Booth
BTW I would NOT include covariates because if the regression ols assumptions are satisfied (in particular, no correlation among IVs(including the covariates) in the path of interest.) the addition of covariates would not change whether or not mediation occurs.
In contrast to David, I would always include all covariables of which I think they are scientifically relevant (even if they don't improve the fit). A model represents aspects of our understanding, and they should be build bason on understanding. Data may then be gathered and fitted to the model to see if there is reason to revise our understanding or to see which of possible alternative models (representing different aspects of our understanding) is better supported by the available data.
As Jochen Wilhelm, I would always include all covariables of which I think they are scientifically relevant (even if they don't improve the fit).
However, if we are in a very exploratory mode and the covariate has little effect on our model, then I probably delete it just for simplicity.
I certainly agree with Roberto: "if we are in a very exploratory mode and the covariate has little effect on our model, then I probably delete it just for simplicity". --
But I have to make a note of caution: many think that "non significant" variables can be omitted, but that's a faulty logic. A variable can be very relevant in a model without being "significant". One really has to to check the consequences: how does including or omitting the variable affect the practical performance of the model?
Or the other way around: a variable can be very "significant" but have only a negligible impact on the model and therefore be removed (if it is not considered scientifically relevant).
Take home: never judge importance or relevance by p-values. Don't do variable selection based on statistical measures like p-values or alikes (R², likelihood, information criteria etc.).
Thank you all! Roberto Molteni , Jochen Wilhelm I, then, have an additional problem. In the step 2 of Baron and Kenny I'm supossed to test the relationship of my independent variable and my mediator, using my mediator as a dependent variable. However, as my study is not focused in this variable, I don't have the theory to include more control variables that may me important for this model. As I'm doing this mediation analysis with 5 other variables, I think It would be inefficient to do complete regression models of each mediator. Can you think of a simpler way out?
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics